Efficient computation of spectral bounds for Hessian matrices on hyperrectangles for global optimization (Q2250109)

Let \(\varphi :U\rightarrow {\mathbb R}\) be a twice continuously differentiable function on an open set \(U\subseteq {\mathbb R}^n\) and let \( B = [\underline x_1, \overline x_1]\times \ldots \times [\underline x _i, \overline x_n ] \) be a closed hyperrectangle in \(U\). The present paper concerns the following problem \[ \begin{aligned}& \text{Find }\underline \lambda\in {\mathbb R}, \overline \lambda \in {\mathbb R} \text{ such that }\\ &\underline \lambda\leq \lambda \leq \overline \lambda \text{ for all eigenvalues } \lambda \text{ of all matrices } H \in H (\varphi, B), \end{aligned} \] where \( H (\varphi, B)\) is the set of Hessian matrices of \(\varphi\) on B, \[ H (\varphi, B) = \{ \nabla^2 \varphi (x): x \in B \}. \] A bound \(\overline \lambda\) (resp. \(\underline \lambda\)) is called tight if there exists at least one matrix \(H\) in the matrix set with an eigenvalue \(\lambda = \overline \lambda\) (resp. \(\lambda= \underline \lambda\)). Note that the bounds \(\underline \lambda, \overline \lambda\) above may or may not be tight. The authors compare two established and a new methods for the calculation of spectral bounds for Hessian matrices on hyperrectangles by applying them to a large collection of 1,522 objective and constraint functions extracted from benchmark global optimization problems. Both the tightness of the spectral bounds and the computational effort of the three methods, which apply to \(C^2\) functions \(\varphi : {\mathbb R}\rightarrow {\mathbb R}\) that can be written as codelists, are assessed.